However, Nvidia’s Arm-based technology appears to be more efficient.
Nvidia is mostly known as a graphics processing unit (GPU) manufacturer; however, the business is also working toward becoming a central processing unit (CPU) manufacturer. Grace is the name of the company’s Arm-based server central processing unit (CPU), which it sells in combinations together with either two Grace CPUs or alongside a Hopper GPU. Both of these products are designed to perform server-related tasks. However, Nvidia is going up against Intel, AMD, and other server CPUs that are based on the Arm architecture, so it is going to have a difficult time competing in this market. Now, the first high-performance computing (HPC) benchmarks for Grace have been published. Although they are a mixed bag for Nvidia’s relatively new Arm CPU, they also show promise. Grace is a high-performance computing (HPC) benchmark.
The benchmarks for Grace have been developed by Thenextplatform and have been obtained from a number of supercomputing labs that have put Grace through its paces throughout the process. Grace has been put through its paces in dual-CPU and CPU+GPU configurations, which are referred to as Grace-Grace and Grace-Hopper respectively, according to the Barcelona Supercomputing Center and the State University of New York. There have been tests conducted on the Nvidia hardware against x86 server CPUs, and the results have been quite fascinating. For your reference, the Grace central processing unit (CPU) from Nvidia has 72 Arm cores and 480GB of LPDDR5X memory. This makes it a unique CPU because it offers a high-core count with two Grace chips (the default) delivering 144 cores. Additionally, these Grace chips use mobile memory (with ECC) rather than HBM, which is the case with the majority of server CPUs.
Nvidia’s Grace products were very competitive with older architectures such as Intel Skylake and AMD Milan, for example, and appeared to have a significant advantage when it comes to cost and thermals due to their Arm architecture. The benchmarks include a wide variety of tests, and the TL;DR is that the benchmarks include a wide variety of tests. However, in some high-performance computing benchmarks, the Grace central processing units (CPUs) from Nvidia were defeated by the Sapphire Rapids CPUs from Intel. Nvidia’s Grace, for instance, was only slower in three out of eight tests when compared against a pair of 48-core Xeon Max 9468 central processing units (CPUs) manufactured by Intel, as stated by Tom’s Hardware.
In general, the findings for Grace are quite outstanding when compared to those of other platforms, and it is possible that it also has a considerable edge in terms of thermals. It should be noted, however, that the benchmarks did not include any data regarding power consumption. A dual-socket Sapphire Rapids chip has a thermal design power (TDP) of 700W, but the dual-Grace chip from Nvidia has a TDP of 500W. This indicates that Nvidia has a significant advantage in this field, which is an essential consideration for the implementation of data centers.
The fact that none of these tests contain AMD’s most recent Zen 4 server chips, which are based on Epyc Genoa central processing units, is the most significant omission. A significant number of data center operators will be looking closely at both AMD and Nvidia in the near future. This is because AMD is beginning to take more market share from Nvidia in the server industry.